[Intel-gfx] [PATCH 1/2] drm/i915: fix possible RPM ref leaking during RPS disabling

[Daniel Vetter]

On Thu, May 22, 2014 at 01:56:17PM +0100, Robert Beckett wrote:
> On 13/05/2014 14:54, Daniel Vetter wrote:
> >On Tue, May 13, 2014 at 04:46:10PM +0300, Imre Deak wrote:
> >>On Mon, 2014-05-12 at 19:51 +0200, Daniel Vetter wrote:
> >>>On Mon, May 12, 2014 at 06:35:04PM +0300, Imre Deak wrote:
> >>>>In
> >>>>
> >>>>commit c6df39b5ea6342323a42edfbeeca0a28c643d7ae
> >>>>Author: Imre Deak <imre.deak at intel.com>
> >>>>Date:   Mon Apr 14 20:24:29 2014 +0300
> >>>>
> >>>>     drm/i915: get a runtime PM ref for the deferred GT powersave enabling
> >>>>
> >>>>I added an RPM get-ref when enabling RPS from a deferred work, but forgot
> >>>>to add the corresponding put-ref when canceling the work. This may leave
> >>>>RPM disabled.
> >>>>
> >>>>Signed-off-by: Imre Deak <imre.deak at intel.com>
> >>>
> >>>Could we intentionally hit this by e.g. racing a suspend/resume against
> >>>runtime pm? E.g.
> >>>1. disable all outputs, make sure we've entered runtime pm
> >>>2. set runtime autosuspend delay to 0
> >>>3. suspend/resume
> >>>4. autosuspend (hopefully, my understanding is a bit unclear)
> >>>
> >>>->Boom
> >>>
> >>>Would look nice as an igt subtest if it works ;-)
> >>
> >>Yep, works consistently as expected both before and after the fix. I
> >>pushed the new subtest, please add here:
> >>Testcase: igt/pm_pc8/system-suspend
> >
> >Now I'm confused: I expected this to blow up without your fix here, and
> >not work with or without it?! Please unconfuse ...
> >-Daniel
> >
> Im not sure what you are expecting to go boom.
> It is pretty difficult to get a race between suspend/resume and runtime pm.
> During system suspend, the pm core grabs a runtime ref without resume
> callbacks, and releases it on the resume. (see
> https://www.kernel.org/doc/Documentation/power/runtime_pm.txt part 6).
> 
> Also, as i915_drm_freeze does a runtime pm get, at step 3, it will cause a
> runtime resume before doing the system suspend. On the system resume, it
> will do a runtime pm put, which will cause a runtime suspend due to
> autosuspend delay being 0, and so step 4, autosuspend will occur. This
> should all work fine (assuming the patch is applied to handle any
> refcounting between back to back resume->suspend which cancels the enable
> before it is executed).

Rest assured that Imre's igt goes indeed boom. We run the rps stuff
delayed which makes the race possible, as long as you autosuspend quickly
enough. You might want to run the testcase to confirm that ;-)

And thanks a lot for mentioning your concerns, this kind of information
and questions are exactly what I expect from patch reviewers. I will
augment the commit message a bit to make it clear how stuff blows up.

> It all looks fine to me.
> 
> Reviewed-by: Robert Beckett <robert.beckett at intel.com>

Thanks, Daniel

> 
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch